[Model] Avoid scalar synchronizations in GLM-Image preparation

[yuweih205]

Purpose

GLM-Image materializes post-patch height and width as CUDA scalar tensors in
both the image projector and prepare module. Using those values for reshaping
causes four device-to-host scalar copies per forward. The sequence-parallel
padding path also reduces a mask that was just given a nonempty false tail,
adding a fifth copy when padding is present.

Keep shape metadata as Python integers or symbolic integers, derive the output
shape from the original latent shape, and return only the three tensors consumed
by the sequence-parallel plan. The padded-mask reduction is removed because it
cannot be true in that branch.

Test Plan

python -m pytest -q \
  tests/diffusion/models/glm_image/test_glm_image_quantization.py \
  -k TestGlmImagePrepareModule

vLLM Version: 0.28.0

vLLM-Omni Commit: 615f97b9c79e2143c91d3f7fe81603327df059aa

Test Result

1 passed, 44 deselected, 16 warnings in 3.30s

The focused test ran in Python 3.12 with the distributed backend set to Gloo
for the CPU environment.

A source-locked component benchmark ran on eight H200 GPUs with
PyTorch 2.13.0+cu130 and CUDA 13.0. Each worker used 10 warmups followed by
seven ABBA rounds of 100 synchronized calls.

Case	Baseline	This PR	Median reduction	Per-GPU range	D2H copies
1024x1024, BF16, SP1	0.4043 ms	0.2562 ms	36.78%	35.91%-37.25%	4 -> 0
992x1056, BF16, simulated SP8 padding 4092 -> 4096	0.4655 ms	0.2865 ms	38.39%	37.64%-38.80%	5 -> 0

All tensor values, shapes, dtypes, devices, layouts, strides, storage offsets,
and contiguity matched exactly on 8/8 GPUs; the candidate was faster on 8/8.
These measurements cover GlmImagePrepare plus outer shape/mask metadata,
not full DiT or end-to-end generation.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/diffusion/diffusion_model_integration.md.

Module owners: @fhfuih @Bounty-hunter @wtomin @RuixiangMa

@yuweih205, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[yuweih205]

Self-review for e4ac6f50: I checked that the image projector now computes post-patch height and width as Python integers, that GlmImagePrepare no longer materializes those scalars and returns only the three tensors consumed by the _sp_plan, and that the outer forward derives them from the original latent shape with the same // patch_size arithmetic instead of .item() calls. The removed hidden_states_mask.all() check sat directly after the padded tail was written to False, so it could not be true in that branch.

The focused TestGlmImagePrepareModule test was updated for the new constructor and 3-tuple and passed on CPU with Gloo: 1 passed, 44 deselected. The source-locked component benchmark on eight H200 GPUs showed 36.78% (SP1) and 38.39% (simulated SP8 padding) median reductions with D2H copies going 4 -> 0 and 5 -> 0, and all tensor values, shapes, dtypes, strides, and layouts matched exactly on 8/8 GPUs.

These measurements cover GlmImagePrepare plus the outer shape/mask metadata, not full DiT or end-to-end generation.

[gcanlin]

LGTM, thanks